Identification of Papilionidae Species in Yunnan Province Based on Deep Learning

Abstract

Yunnan is known as the "Hometown of Butterflies" in China. The colorful and morphological diversity of the Papilioidae in Yunnan province is the subject of insect ecology and evolution research. At the same time, the Yunnan Papilio has great ornamental value. It is of great significance to accurately identify the species of Papilionidae in Yunnan Province. At present, Yunnan Papilio has not been classified in the related research on butterfly identification using deep learning methods, and there is a situation that the sample data set between species is small and the number is unbalanced, which may cause the model to fail to learn the morphological characteristics of butterflies. In response to the above problems, this study established a data set consisting of 12,956 original images of papilionidae from Yunnan Province, including two subfamilies, 12 genera and 80 species. Five deep learning network models (VGG-19, ResNet-34, ResNet-50, ResNet-101 and DenseNet-121) were explored from the perspective of prediction accuracy and loss value by transfer learning method. And modeling effects of SGD, Adam, Adamax and RMsprop optimization algorithms. The final data set adopts balanced sampling and 11 data enhancement methods for data fusion to expand the data set to 16,000 images. The ResNet-50 network structure optimized by Adamax algorithm is selected to achieve the optimal effect. The experimental results show that the recognition accuracy of ResNet-50 in the constructed model reaches 87.47%. The study provides a basis for constructing a visual recognition model of Papilioidae in Yunnan and applying it to the mobile terminal, and provides a fast and efficient new method for species identification of Papillidae in Yunnan. (Abstract)